MODIFIABLE AREAL

UNIT PROBLEM

S. Openshaw

ISSN 0306-6142
ISBN 0 86094 134 5
S. Openshaw
Published by Geo Books. NorwichPrinted by Headley Brothers Ltd. Kent
CATMOG - Concepts and Techniques in Modern Geography
CATMOG has been created to fill in a teaching need in the field of quantitative CONCEPTS AND TECHNIQUES IN MODERN GEOGRAPHY No.38
methods in undergraduate geography courses. These texts are admirable guides for
teachers, yet cheap enough for student purchase as the basis of classwork. Each
book is written by an author currently working with the technique or concept he
describes. THE MODIFIABLE AREAL UNIT PROBLEM

I. Introduction to Markov chain analysis - L. Collins

2. Distance decay in spatial interactions - P.J. Taylor by
3. Understanding canonical correlation analysis - D. Clark
4. Some theoretical and applied aspects of spatial interaction shopping models Stan Openshaw
- S. Openshaw
5. An introduction to trend surface analysis - D. Unwin (Newcastle University)
6. Classification in geography - R.J. Johnston
7. An introduction to factor analysis - J.B. Goddard & A. Kirby
8. Principal components analysis - S. Daultrey
9. Causal inferences from dichotomous variables - N. Davidson
CONTENTS
10. Introduction to the use of logit models in geography - N. Wrigley
II. Linear programming: elementary geographical applications of the transportation page
problem - A. Hay 3
An introduction to quadrat analysis (2nd edition) - R.W. Thomas
I INTRODUCTION
1 2.
13. An introduction to time-geography - N.J. Thrift
II AN INSOLUBLE PROBLEM OR A POTENTIALLY POWERFUL GEOGRAPHICAL TOOL?
14. An introduction to graph theoretical methods in geography - K.J. Tinkler
15. Linear regression in geography - R. Ferguson 7
16. Probability surface mapping. An introduction with examples and FORTRAN
(i) An insoluble problem
programs - N. Wrigley
(ii) A problem that can be assumed away 7
17. Sampling methods for geographical research - C.J. Dixon & B. Leach
18. Questionnaires and interviews in geographical research - C.J. Dixon & B. Leach 7
19. Analysis of frequency distributions - V. Gardiner & G. Gardiner (iii) A powerful analytical device
20. Analysis of covariance and comparison of regression lines - J. Silk
21. An introduction to the use of simultaneous-equation regression analysis in
geography - D. Todd III ON THE NATURE OF THE MODIFIABLE AREAL UNIT PROBLEM
22. Transfer function modelling: relationship between time series variables
(i) Definitions 8
- Pong-wai Lai
23. Stochastic processes in one dimensional series: an introduction - K.S. Richards
(ii) Early correlation studies 10
24. Linear programming: the Simplex method with geographical applications
- James E. Killen
(iii) More recent studies 13
25. Directional statistics - G.L. Guile & J.E. Burt
26. Potential models in human geography - D.C. Rich
27. Causal modelling: the Simon-Blalock approach - D.G. Pringle
28. Statistical forecasting - R.J. Bennett IV THE RESULTS OF SOME AGGREGATION EXPERIMENTS
29. The British Census - J.C. Dewdney
The analysis of variance - J. Silk (i) Random aggregation and the correlation coefficient 16
30.
31. Information statistics in geography - R.W. Thomas
(ii) Random aggregation and other statistics 19
32. Centrographic measures in geography - A. Kellerman
33. An introduction to dimensional analysis for geographers - R. Haynes
34. An introduction to Q-analysis - J. Beaumont & A. Gatrell (iii) Random aggregation experiments with once aggregated data 19
35. The agricultural census - United Kingdom and United States - G. Clark
- G. Aplin (iv) Identifying the limits of the MAUP 21
36. Order-neighbour analysis
37. Classification using information statistics - R.J. Johnston & R.K. Semple
38. The modifiable areal unit problem - S. Openshaw (v) Spatial calibration of a statistical model 25
39. Survey research in underdeveloped countries - C.J. Dixon & B.E. Leach
(vi) .. but do the optimal zoning systems look nice? 30
This series is produced by the Study Group in Quantitative methods, of the Institute
of British Geographers.
For details of membership of the Study Group, write to the Institute of British
Geographers, I Kensington Gore, London SW7 2AR, England.

The series is published by: 1

Geo Books, Regency House, 34 Duke Street, Norwich NR3 3AP, England, to whom
all other enquiries should be addressed.
 V POSSIBLE SOLUTIONS

(i) No philosopher's stone 31 I INTRODUCTION

(ii) Non-geographical solutions 32 The usefulness of many forms of spatial study, quantitative or other-
wise, depends on the nature and intrinsic meaningfulness of the objects that
(iii) A traditional geographical solution 33 are under study. Geographers have a long tradition of studying data for
areal units; for example, spatial objects such as zones or places or towns
(iv) Towards a new methodology for spatial study 34 or regions. The problem is that ever since the demise of 'the region' as
the primary object of geographical study very little concern has been ex-
pressed about the nature and definition of the spatial objects under study.
VI CONCLUSIONS 37 As Chapman (1977) put it 'Geography has consistently and dismally failed to
tackle its entitation problems, and in that more than anything else lies the
root of so many of its problems' (page 7). In short insufficient thought is
BIBLIOGRAPHY 39 given to precisely what it is that is being studied.

ACKNOWLEDGEMENTS
Particular thanks are due to Dr A.C. Gatrell, Dr K. Jones, and two
anonymous referees for making many useful and extremely helpful
suggestions.
For many purposes the zones in a zoning system constitute the objects
or geographical individuals that are the basic units for the observation and
measurement of spatial phenomena. It is usual in a scientific experiment that
the definition of the objects of study should precede any attempts to measure
their characteristics. However, this is not the case with areal data where
the spatial objects only exist after data collected for one set of entities
are subjected to an arbitrary aggregation to produce a set of spatial units.
Consider an example about wheat and potato yields. Data for one set of
entities (farms or fields) can be aggregated to produce data for a set of
spatial entities (parishes or counties). In this instance spatial aggrega-
tion is necessary in order to 'create' a relevant data set. As Yule and
Kendall (1950) put it '.. geographical areas chosen for the calculation of
crop yields are modifiable units, and necessarily so. Since it is impossible
(or at any rate agriculturally impracticable) to grow wheat and potatoes on
the same piece of ground simultaneously we must, to give our investigation
any meaning, consider an area containing both wheat and potatoes and this
area is modifiable at choice' (page 312). What they mean is that it is
necessary to use areal units that are larger than the individual field and
include both wheat and potatoes so that some measure of spatial association
can be computed. Obviously at the level of the individual field there is no
spatial association (assuming the fields are either all wheat or all potatoes)
and, therefore, the degree of spatial association depends on the nature of
the areal units that are used. The definition of these geographical objects
is arbitrary and (in theory) modifiable at choice; indeed, different research-
ers may well use different sets of units. This process of defining or cre-
ating areal units would be quite acceptable if it were performed using a
fixed set of rules, or so that there was some explicit geographically meaning-
ful basis for them. However, there are no rules for areal aggregation, no
standards, and no international conventions to guide the spatial aggregation
process. Quite simply, the areal units (zonal objects) used in many geo-
graphical studies are arbitrary, modifiable, and subject to the whims and
fancies of whoever is doing, or did, the aggregating. It is most unfortun-
ate that there is no standard set of spatial units.

Since any study region over which data are collected is continuous, it
follows that there will be a tremendously large number of different ways by
which it can be divided into non-overlapping areal units for the purpose of

2 3
spatial analysis. Viewed as a combinatorial problem, the number of different
zoning systems, each of m-zones, to which data for n individuals can be ag-
gregated becomes incredibly large even for small values of n (Keane, 1975).
For example, there are approximately 10 12 " different aggregations of 1,000
objects into 20 groups. If the aggregation process is constrained so that
the groups consist of internally contiguous objects (i.e. all the objects
assigned to the same group are geographical neighbours) then this huge number
is reduced, but only by a few orders of magnitude. So even with the im-
position of contiguity constraints the combinatorial problem remains totally
unmanageable.
Consider an example based on census data. In Tyne and Wear County there
are about 1.1 million people and 300,000 households. The 1981 census uses
a set of about 2,800 enumeration districts to report the results. Consider
how many different sets of 2,800 zones could be used for reporting the census
characteristics of 300,000 households: Moreover, there are other huge com-
binatorial explosions whenever a zoning system of 2,800 zones are re-aggregated
to form other zoning systems with fewer zones; for example, the 258 zones used
for transportation modelling and planning. There are a tremendously large
number of alternative 258 zone aggregations that could be used, most (if not
all) of which will yield different results.
This, then, is the crux of the modifiable areal unit problem (MAUP).
There are a large number of different spatial objects that can be defined and
few, if any, sets of non-modifiable units. Whereas census data are collected
for essentially non-modifiable entities (people, households) they are reported
for arbitrary and modifiable areal units (enumeration districts, wards, local
authorities). The principal criteria used in the definition of these units
are the operational requirements of the census, local political considera-
tions, and government administration. As a result none of these census areas
have any intrinsic geographical meaning. Yet it is possible, indeed very
likely, that the results of any subsequent analyses depend on these defini-
tions. If the areal units or zones are arbitrary and modifiable, then the
value of any work based upon them must be in some doubt and may not possess
any validity independent of the units which are being studied.
The question is, does it matter? If you change the areal basis does it
have any really significant effect on the results? Do haphazard zoning
systems yield haphazard results? If they do, then what can be done about it?
Consider two more examples. The definition of enterprise zones was
restricted to areas with high levels of unemployment. Unemployment rates were
calculated for a set of statistical reporting units known as 'travel to work
areas' (TTWAs); for details see Coombes and Openshaw (1982). Unfortunately,
for a few areas of the country these particular areal units provide a poor
representation of labour markets and present a biased picture of levels of
unemployment. For example, for some obscure reason South Tyneside (in Tyne
and Wear County) was included in the same TTWA as Washington, Gateshead,
Jarrow, and parts of rural Northumberland. The effect was to mix areas of
very high unemployment with fairly prosperous rural areas which have no strong
journey to work links; the result was to reduce the apparent level of un-
employment on South Tyneside. The total June unemployment, rates for the
period 1978-82 were 11.1, 10.7, 12.9, 17.2, 18.7. However, if a more geo-
graphically meaningful definition of the South Tyneside labour market is used
(see Coombes et al, 1982) then the unemployment rates for South Tyneside be-
come 13.9, 13.3, 15.9, 20.1, 20.7; more than enough to justify an enterprise
4 5
zone. The unfortunate use of the 'wrong' set of areal units has therefore
deprived South Tyneside of considerable job opportunities.
A final example concerns the Parlimentary Boundary Commision which
reviews the boundaries of the 520 English constituencies every 15 years. This
is a pure exercise in modifying areal units. The task is to select an
appropriate amalgamation of wards to create constituencies of approximately
equal electoral population size, that conform as far as practicable to county
boundaries and London Boroughs, and that take into account local community
interests. The latest revisions (1983) were performed by manual means and
have been heavily criticised because of inconsistencies in application and
unequal constituency sizes. The average electorate is about 68,000 but it
varies from extremes of 24,000 (Newcastle Central) to 100,000 (Buckingham).
There are even larger discrepancies in neighbouring constituencies; 57,000 in
Finchley but 84,000 in Wood Green. The reason is simply that only a small
proportion of all alternative areal arrangements were identified (Johnston
and Rossiter, 1983). For example, the 26 wards in Camden can be aggregated
to form two constituencies in 878 different ways, with a maximum two percent
deviation from the mean size of 68,000 (Johnston, The Times, March 15th 1983).
If a political geographer had ward-level voting data, then these different
constituency definitions would yield a wide range of results.
If the reader is still unconvinced then he can attempt the following
experiment. Construct an artificial set of areal units, or use a map of a
few neighbouring local authorities. Assign some data to them. Compute either
a few statistics for each zone (eg rates) or an overall statistic (eg cor-
relation coefficient or mean). Now amalgamate a few zones which are contig-
uous, re-calculate your statistics for the aggregated data, and examine the
changes. Now try to amalgamate a few more zones with the aim of either
increasing or reducing the magnitude of the changes. Obviously this experi-
ment would be easier if a microcomputer was used. However, a few hours exp-
erimentation will convince virtually anyone about the severity of the MAUP.
Quite simply, different aggregations yield different results but without any
systematic trends emerging that can be used for prediction or correction
purposes.
What is so surising about the MAUP is that while geographers know of
its existence they readily assume, in the absence of any knowledge, that it
has no significant effect on their studies. An important reason for this
deliberate neglect is that the validity of many applications of quantitive
analyses of zonal data depends on the assumption that the MAUP does not exist
and that the spatial units under study are given, meaningful, and fixed.
Whilst these may be tolerable assumptions for a statistician, who may know
no better, it is hardly a satisfactory basis for the application and fruther
development of spatial analysis techniques in geography (Openshaw and Taylor
1981).
Although there is an almost infinite number of different ways by which
a geographical region of interest can be areally divided, data are normally
only presented and analysed for one particular set of units. The choice of
these units is often haphazard, in that considerations such as convenience
rather than geographical meaning are paramount. This uncertainty about the
nature and definition of the zonal objects of spatial study is an important
consequence of the MAUP. It is important because of the effects that the use
of different areal units may have on the results of geographical study and
because it is endemic to all analyses of areal or zonal data. It is a major
geographical problem with ramifications that need to be properly appreciated
by geographers and all others interested in the analysis of spatially aggre-
gated data. Looked at in this way, the MAUP is today one of the most important
unresolved problems left in spatial analysis. There has been very little
research compared with that afforded to many far less significant problems,
and whilst it appears to be primarily a technical problem, it is also a major
conceptual problem that is central to many aspects of geographical study.
II AN INSOLUBLE PROBLEM OR A POTENTIALLY POWERFUL GEOGRAPHICAL TOOL?
Before examining various empirical evidence about the nature and severity
of the MAUP, it is useful to consider further some of the more general aspects
of the problem. This is important because the viewpoint that one adopts
heavily influences thinking about how best to handle it.
It is unfortunate that so many geographers have become so completely
blinkered by the concepts of conventional statistical theory and the normal
science paradigm that they no longer seem to care or understand the basic
geography of what they are doing. A decade or so ago this could be readily
justified by the importance of introducing scientific methods and quantitative
techniques into geography; the hope being expressed that the various un-
realistic statistical and geographical assumptions could be relaxed later.
To some extent this has happened and the significance of many purely statis-
tical problems caused by the peculiar nature of areal data has been reduced;
for example, by the development of methods capable of handling spatially
autocorrelated data (Cliff and Ord, 1975). However, many of these statistical
advances have been made at the expense of geographical considerations. It
has perhaps been quietly overlooked that statistical techniques still cannot
cope with the modifiable nature of areal data. The attempts by Griffith (1980)
and others to establish a theory of spatial statistics are doomed to fail for
the simple reason they begin with the assumption that the data under study
are fixed!
It is humbly suggested that it is about time that quantitative geo-
graphers started to devise a body of relevant spatial analysis techniques
that can cope with geographical data. The first stage in this second quanti-
tative revolution must be the development of methods that can handle the MAUP,
with the gradual replacement of many of the less relevant techniques that
were originally plagiarised from a variety of disciplines in the 1960's and
1970's. If geography is to survive as a distinctive subject then it is time
it stopped copying and adapting techniques imported from other disciplines
and started a period of fundamentally relevant methodological innovation.
There may well have to be a move away from a rather rigid and naive approach
based on classical statistical theory.
The MAUP is sufficiently important that it presents a good opportunity
for the development of new geographical techniques. However, before this
can happen it is necessary to adopt a realistic perspective as to the im-
portance of the MAUP to geography. To assist this process, it is useful to
briefly view the modifiable areal unit problem from three different per-
spectives.
6 7
(i) An insoluble problem
One very good reason for ignoring the MAUP is the belief that it is in-
soluble. If it really is endemic to the study of all areal data and if it
really is insoluble then why not pretend it does not exist, in order to allow
some analysis to be performed? This CATMOG is dedicated to those who believe
in this fallacy of insolubility.
(ii) A problem that can be assumed away
It is always possible, at least in theory, to change the nature of the
MAUP by assuming it away. For example, a zonal model of trip rates made by
members of a household can be re-expressed as a disaggregate model at the
i ndividual level, although there may be some problems if the zonal variables
do not exist at the individual level. Other problems relate to the difficult
statistical problems of identifying the nature of the underlying model im-
plicit in aggregate level study and of estimating its unknown parameters if
only aggregate data are available. It is also apparent that this approach
is not always applicable; for instance, the crop yield example quoted from
Yule and Kendall (1950). Even more important from a geographical point of
view, it amounts to trying to write out or assume away any spatial effects
and is, therefore, intrinsically non-geographical and aspatial; good statis-
tics but very poor geography. It is of little use if the purpose of spatial
study is to investigate spatial associations and it is argued that it is
precisely this interest in areal phenomena that is a unique characteristic
of geographical study.
(iii) A very powerful analytical device
If the MAUP is endemic to spatial study and if it cannot simply be ig-
nored or assumed away then methods should be developed to handle it and bring
it under control in a purposeful way. Thus it can be regarded both as an
opportunity for geographers to develop a new approach to spatial study based
on zonal data and as a potentially very powerful geographical tool once the
tremendous aggregational uncertainty or spatial freedom inherent in the MAUP
can be usefully exploited for geographical purposes. Thus the MAUP is only
a problem if it is viewed from a perspective that cannot handle it. It is
certainly true that it is unlikely to have a precise analytical solution.
However, the availability of fast super-computers opens up the possibility
of seeking approximate numerical solutions. If the MAUP can be manipulated
to suit particular purposes, via a kind of spatial optimisation process, then
is it not possible that the resulting optimal zoning systems can be used for
a number of geographical purposes and as a basis for a new approach to spatial
study?
The suggestion made here is that this third perspective of the MAUP is
the most appropriate one for geographers. It is, after all, a geographical
problem that requires a geographical rather than a statistical solution.
This CATMOG argues in favour of this view. It does so by first describing
the nature and historical background of the MAUP followed by the results of
a series of simple experiments involving nothing more complex than a cor-
relation coefficient. Finally, two alternative geographical solutions are
described with some discussion of areas where further developments may be
expected in the future.
 III ON THE NATURE OF THE MODIFIABLE AREAL UNIT PROBLEM

(i) Definitions

The MAUP is in reality composed of two separate but closely related

problems. The first of these is the well known scale problem which is the
variation in results that can often be obtained when data for one set of areal
units are progressively aggregated into fewer and larger units for analysis.
For example, when census enumeration districts are aggregated into wards,
Districts, and Counties the results change with increasing scale. Previously,
geographers have been very interested in scale problems of this sort, largely
because it was thought that systematic scale effects could be easily handled.

Although scale differences are a most obvious manifestation of the MAUP

there is also the problem of alternative combinations of areal units at equal
or similar scales. Any variation in results due to the use of alternative
units of analysis when the number of units is held constant is termed the
aggregation problem (Openshaw, 1977a).

The MAUP obviously includes both these subproblems. The scale problem
arises because of uncertainty about the number of zones needed for a particu-
lar study. The aggregation problem arises because of uncertainty about how
the data are to be aggregated to form a given number of zones. It should be
noted that for any reasonably sized data set there is considerably more
spatial freedom in the choice of aggregation than there is in the choice of
the number of zones.

At this stage it is worth noting that there are two different types of
zonal arrangement. Most geographical studies have employed spatial aggrega-
tions based on contiguous arrangements of zones, something referred to as a
zoning system. However, a zoning system is only a special case of a grouping
system that incorporates a contiguity constraint. The non-contiguous case
is referred to as a grouping system. The use of a contiguity constraint
restricts the degree of aggregational variability but inmost practical studies
it is so large anyway that it brings little real advantage other than the
convenience of having zones which are formed of internally connected units.

Finally, it is noted that the MAUP is also closely involved in what is

known as the ecological fallacy problem . An ecological fallacy occurs when
it is inferred that results based on aggregate zonal (or grouped) data can
be applied to the individuals who form the zones or groups being studied.
In a geographical context the individuals can either be zones prior to a
subsequent aggregation or non-modifiable entities. Obviously whether the
ecological fallacy problem exists or not depends on the nature of the aggre-
gation being used. A completely homogeneous zoning or grouping system would
be free of this problem. However, most, if not all, zoning systems studied
by geographers are internally heterogeneous so that the severity of any eco-
logical fallacy problem depends largely on the nature of the aggregation
being studied.

Figure 1 shows some simple examples of scale and aggregation problems.

The reader is invited to investigate the effects of scale and aggregation by
assigning some arbitrary values to the zones, computing a correlation
Figure 1. Alternative aggregations of 16 zones into 8 and 4 regions
8 9
coefficient (for instance) with a calculator and then repeating the process
for different data aggregations. The data may be aggregated either by aver-
aging or by adding rates, or by complete recalculation of numerators and
denominators. The reader can also examine these effects.
(ii) Early correlation studies
One of the first papers to consider this type of problem is that by
Gehlke and Biehl (1934). They observed that the size of the correlation
coefficient increased with aggregation. The 252 census tracts in Cleveland,
USA, were grouped successively into larger units of approximately the same
size and subject to contiguity restrictions. The correlation between male
juvenile delinquency and median monthly income was then calculated using both
absolute numbers and ratios; see Table 1.
Table 1. Correlation coefficients for Cleveland, USA
Does it measure the inter-relation of traits in their ultimate possessors -
individuals and families? A relatively high correlation might conceivably
occur by census tracts when the traits so studied were completely dissociated
in the individuals or families of those traits' (page 170). Finally, they
asked what is probably the most important question of all concerning whether
a geographical area is an entity possessing traits or merely one character-
istic of a trait itself? That is to say, are areal units entities or objects
that can be studied or are they merely a variable that is proxy for geo-
graphical location?
Yule and Kendall (1950) added to Gehlke and Biehl's findings, demon-
strating in particular that the correlation coefficient usually tends to in-
crease with scale. They describe how the correlations between wheat yields
and potato yields for the 48 counties of England increase as spatial aggre-
gation reduces the number of areal units and increases their size and the
scale of the analysis; see Table 2.

Table 2. Correlations between wheat and potato yield (English counties)

number of units absolute numbers rates
number of geographical areas correlation
252 -.502 -.516
200 -.569 -.504 48 .2189
175 -.580 -.480 24 .2963
150 -.606 -.475 12 .5757
1 25 -.662 -.563 6 .7649
100 -.667 -.524 3 .9902
50 -.685 -.579
25 -.763 -.621 They note that we seem able to produce any value of the correlation from
0 to 1 merely by choosing an appropriate size of the unit of area for which
The principal effect of using ratio variables is to slow down the increase we measure the yields. Is there any "real" correlation between wheat and
of the correlation coefficient due to increasing scale by standardising for potato yields or are our results illusory?' (page 311). There is little
the size of areas. Gehlke and Biehl (1934) then compared these results with doubt that Yule and Kendall had a deep appreciation of the importance of MAUP.
some random groupings of the data without contiguity restrictions; these They recognised the difference that exists between studies based on modifiable
aggregations produced correlations of -.434 for 150 zones and -.544 for 25. units, such as areal units, and those based on non-modifiable units, such as
Random data aggregations have no systematic effect on the correlations. the cow or the shell. Furthermore, they emphasised that in studies based on
modifiable units the magnitude of a correlation will depend on the units that
A second set of experiments was used to demonstrate that the variation are used. In this vein they wrote: 'Our correlations will accordingly measure
in the size of the correlation coefficient was related to the size of the the relationship between the variates for the specified units chosen for the
units involved; the smallest values being associated with the smallest units. work. They have no absolute validity independently of these units, but are
Data from the 1910 census provided two variables (the value of farm products relative to them. They measure, as it were, not only the variations of the
and the number of farmers) for 1,000 rural counties. These data were then quantities under consideration, but the properties of the unit-mesh which we
randomly grouped by Gehlke and Biehl to yield 63 and 31 groups with the fol- have imposed on the system in order to measure it' (page 312). This is a
lowing correlations: very clear early statement of the nature and importance of the MAUP.
n r
1000 .649 Despite this, the correlation coefficient was still regarded as useful
63 .859 because the value for the 48 English counties in 1936 is a geographical and
31 .756 historical fact. A comparison of values for the same units over time might
also be interesting. However, Yule and Kendall consider the result to be
The data were also aggregated to 40 states and 8 counties (zoning systems) specific to the zoning system they used and that it is, therefore, not cap-
with the following results: able of scientific generalisation or for comparison with other correlations
n r
for the same variables but for different zones.
40 .725
8 .826 A feature of both these early studies was the observation that because
Gehlke and Biehl conclude that 'these results raise the question whether a the correlations are modifiable they may not provide any useful guide to
correlation coefficient in census tract data has any value for causal analysis. individual or more spatially disaggregated levels of correlations. Robinson

11
10
(1950) provides the conclusive proof that this in fact the case. He quotes other. This is the interpretation put forward by Taylor (1977). It has also
an example based on the correlation between percentage population 10 years been argued that if the variables are not spatially autocorrelated then the
old and over which is negro and the percentage of the same population that correlation coefficient will not increase with scale. A problem with this
is illiterate; another example is based on the correlation between nativity interpretation is that Blalock ignores aggregation effects and these may
and illiteracy. Table 3 shows the various correlations that were computed easily dominate any scale effects. Some of the ideas developed by Blalock
for different levels of spatial aggregation. (1964) and put into a geographical context by Taylor (1977) have been tested
in Openshaw and Taylor (1979). The expected systematic relationships did
Table 3. Individual and ecological correlations (after Robinson, 1950) not emerge. The effects of the aggregational variability were simply too
strong; indeed, perhaps rather alarmingly, the authors concluded that 'We
level of number correlations between: have been able to find a wide range of correlations. We simply do not know
aggregation of units negroandilliteracy nativity and illiteracy why we have found them. Hence we can make no general statements about vari-
ations in correlation coefficients so that each areal unit problem must be
individual 98 million .203 .118 treated individually for any specific.piece of research' (Openshaw and Taylor,
state 48 .773 -.526 1979; p 142-143). What is meant is that the aggregational variability is not
census division 9 .946 -.619 susceptible to a statistical approach since no systematic empirical regulari-
ties could be found.
The results are quite conclusive. There is a pronounced scale effect in that
the absolute values of the correlations increase as the number of observations (iii) More recent studies
decrease. In addition, the aggregate values bear little resemblance to the
individual values prior to spatial aggregation. Robinson concludes therefore Apart from the occasional mention, the MAUP seems to have been ignored
'..there need be no correspondence between the individual correlation and the until the problem was re-examined in the late 1970's. Openshaw (1977a) was
ecological correlation' (page 354). This is an important result which readily one of the first to re-emphasise the importance of aggregation effects. An
illustrates the dangers of making individual level inferences from analyses example readily shows the importance of the aggregation problem and relative
performed at an aggregate level. insignificance of the scale problem. The data used here relate to 100 metre
grid-squares for South Shields. These data could be readily aggregated to
A final paper that is of interest in this section is that of Blalock 200, 300, 400, 500, 600, 700, 800, 900 and 1 km squares. For each of these
(1964). He describes the results of a series of experiments designed to in- scales there are a number of alternative aggregations; for example, shifting
vestigate the effects of data aggregation. The correlation coefficient be- the origin of the 100 metre lattice produces 25 different 500 metre grid-
tween differences in income for blacks and whites and percentage blacks for square aggregations. The resulting distribution of correlation coefficients
150 southern USA counties was found to be 0.54. Blalock was interested in are shown in Table 5.
the question of what happens if the counties are grouped into larger units in
various different ways. The results are shown in Table 4. Table 5. Scale and aggregation effects on the correlation between
numbers of arly and Mid-Victorian houses in South Shields
Table 4. Blalock's aggregation experiments
size of squares scale aggregation effects
number of units random grouping random zoning ( metres) effects mean correlation standard deviation
75 .67 .63
100 .08 - -
30 .61 .70 .21 .31 .11
200
15 .62 .84 .43 .43 .06
300
10 .26 .81 .47 .11
400 .28
500 .55 .49 .16
With random grouping systems we would expect the correlation coefficients to .52 .16
600 .45
show no systematic scale effects. The variability in values will be due to
700 .20 .57 .18
sampling fluctuation. In this instance sampling fluctuation is in fact the .56 .58 .18
800
aggregation component since there are a very large number of ways by which .66 .60 .19
900
150 objects can be randomly grouped into 75 groups or less. The apparently .62 .20
1 km .73
anomalous value of the correlation coefficient for the 10 groups is an indi-
cation of this effect; indeed, it is slightly miraculous that the other The second column shows the effects of increasing scale using only one of the
values are so uniform. possible aggregations to each scale of grid-square. The third column shows
the mean correlation based on different aggregations to the same scale pro-
By contrast the random zoning systems will be affected by any spatial duced by moving the origin of the lattice. The fourth column shows the
autocorrelation present in the data, so that the rising correlations with standard deviation of the correlation coefficients produced for the different
increasing scale can be regarded as the result of spatial autocorrelation aggregations to each scale. This example contradicts the claim by Evans (1981)
whereby the zoning system retains more variance of one variable than of the that with grid-square data the changes in correlation coefficient are usually

12 13
consistent across a wide range of scales up to 256 km (page 55). The reason Table 6. Cross-tabulation of individual and ecological correlations
is that his 1 km squares have already smoothed the data dramatically. Fin- (percentage of row totals)
ally, it is noted that the example shown in Table 5 does not consider the
full extent of the aggregation problem. This would involve an examination of
10,000 alternative 100 metre squares (assuming the data being aggregated have areal correlations
been grid-referenced at the 1 metre level), 1,000,000 different 1 km squares, individual -1. -.8 -.6 -.4 -.2 .0 .2 .4 .6 .8
and even larger numbers of alternatives if the zones are not constrained to correlations -.8 -.6 -.4 -.2 .0 .2 .4 .6 .8 1. total
be square in shape.

The conclusion that can be drawn from Table 5 is that Yule and Kendall -1. to -.8
(1950) were quite correct, although they clearly underestimated the severity -.8 to -.6
of the problem. Different variables can be affected by aggregation in dif- -.6 to -.4
ferent ways so that multivariate techniques based on correlations will tend -.4 to -.2
-.2 to .0
to amplify the differences in results caused by the use of different zoning
.0 to .2
systems. As a result, the aggregation and scale variability reported for the
.2 to .4
correlation coefficient also applies to more complex multivariate methods and
.4 to .6
to many other forms of analysis. It is demonstrated later that it is not a
.6 to .8
problem that afflicts only the poor correlation coefficient.
totals
The ecological fallacy problem has also been studied further. The
principal problem here is that a detailed investigation requires access to
Sunderland 1 km squares
large spatially referenced individual data sets and it is only quite recently
that sufficiently powerful computers have become available_ to handle these.
The ecological fallacy problem occurs because areal studies cannot distin- -1. to -.8
guish between spatial associations created by the aggregation of data and real -.8 to -.6
associations possessed by the individual data prior to spatial aggregation. -.6 to -.4
Thus the characteristics of typical deprived urban areas need not be the same -.4 to -.2
as the characteristics of the individuals who live there. -.2 to .0
.0 to .2
One consequence of Robinson's work was that may social scientists inter- .2 to .4
preted his warning as a rigid taboo on the use of all aggregate data; although .4 to .6
this never extended to geography. Borgatta and Jackson (1980) pointed out .6 to .8
that 'what happened was the assumption that, because use of aggregate data
could be misleading at the individual level, every such interpretation had to totals
be incorrect' (page 8). It is also possible that Robinson exaggerated the
i mportance of the problem; in particular he only examined the most gross Sunderland polling districts
levels of aggregation. The question arises as to whether these results are
typical of what might happen with finer spatial scales and more realistic
-1. to -.8
zoning systems.
-.8 to -.6
-.6 to -.4
Recently, some further insights into this problem have come from the
-.4 to -.2
analysis of a random 10 per cent sample survey of all households in Sunder-
-.2 to .0
land and from the analysis of individual census data for part of Italy
.0 to .2
(Openshaw, 1983; Bianchi et al, 1981). A brief description of the results
.2 to .4
for Sunderland can best be examined here. These data can be studied at the
.4 to .6
individual level (8,483 households) or aggregated to polling districts (36
.6 to .8
zones), 1 km squares (117 zones), and 500 metre squares (348 zones). A set
of 54 typical indicator variables were computed. The simplest way to in-
totals
vestigate the ecological fallacy problem is to cross-tabulate the individual
and zonal correlation coefficients (Table 6).
Sunderland 500 m squares

The size of the percentages in the diagonals gives an indication of the

extent to which aggregation has either increased or decreased the magnitudes
of the correlation coefficients. A comparison of the row and column totals

14 15
shows that aggregation has a flattening effect on the frequency distribution
of the individual correlation coefficients. Table 6 clearly demonstrates
the systematic biasing of the ecological correlations from 0 towards 1 and
that the magnitude of the bias increases with scale.
It is noted that the cross-tabulations in Table 6 do not give any indi-
cation of aggregational variability since only one aggregation at each scale
was examined. That is to say, these results refer only to scale effects and
it may be expected that the aggregational effects will be somewhat larger.
Both are important since if these phenomena were better understood it might
be possible to design improved areal definitions for reporting census data.
For instance, is there a critical size for census enumeration districts which
may minimise the effects of scale and aggregation on the data being aggregated?
The present size is merely a reflection of the area that can be covered by a
census enumerator in one day; this is hardly a meaningful variable in urban
geography. It is something of a mystery why census data collecting agencies
do not bother to try and resolve these very important practical questions.
These results suggest that perhaps the magnitude of the ecological fal-
lacy problem is less than the results presented by Robinson (1950) might
indicate. Certainly the changes in the magnitude of the correlation coeffi-
cient are smaller in Table 6 than in Table 3. However, this is slightly mis-
leading since only a small percentage of all correlations in Table 6 do not
have substantial and systematic biases; for the polling district data the
figure is 16 per cent. Additionally, it is impossible to predict the severity
of the problem without access to individual data. As a result there is no
way of knowing whether a particular areal data set will yield values which
are close to the individual values. A fuller discussion of empirical aspects
is provided in Openshaw (1983), while Williams (1976, 1979) outlines a
theoretical interpretation.
connections and partly because it has been suggested that the statistical
distribution of a statistic due to sampling variability and its zoning dis-
tribution due to the choice of different zoning systems, are analogous. If
the analogy could be proven then it would be exceptionally convenient because
it would allow the standard formulae for estimating sampling errors for
simple random samples to be used to provide estimates of aggregational vari-
ability, presumably under the assumption of simple random zoning. In this
sampling-zoning analogy the number of zones in the zoning system would be
regarded as equivalent to sample size.
For this study 1970 census data for the 99 counties in the State of Iowa,
USA, are examined. Two variables are selected for analysis; the percentage
vote for Republican candidates in the congressional election of 1968 and the
percentage of the population over 60 years. There is nothing special about
the selection of this data, it merely happened to be convenient! Openshaw
and Taylor (1979) report a range of different correlations that can be pro-
duced for these variables when the 99 counties are aggregated into a number of
arbitrary six zone aggregations; the values ranged from .26 for the con-
gressional districts to .86 for a simple typology of Iowa into rural-urban
types. The value of the correlation at the 99 county level is 0.34. Since
the 99 counties form a complete population of Iowa counties this value can
be regarded as the population correlation. The question is how well random
samples and sample random zoning systems represent this population value.
Table 7 reports the means and standard deviations of the correlation
coefficient for 10,000 random samples of (i) random zoning systems (randomly
selected areal aggregations) with 6, 12, 18, 24, 30, 36, 42, 48, and 54
zones; and (ii) random samples (random selections of various numbers of zones)
of 6, 12, 18, 24, 30, 36, 42, 48, and 54 counties. The latter provide re-
sults which approximate the values that would be obtained from standard samp-
ling formulae. Openshaw (1977b) describes the computer algorithm used to
generate the quasi-random zoning systems.

IV THE RESULTS OF SOME AGGREGATION EXPERIMENTS Table 7. Sampling and zoning distributions of the correlation coefficient

(i) Random aggregation and the correlation coefficient number zoning distributions
of zones mean standard deviation sample size mean standard deviation
The complex nature of the MAUP suggests that further advances in our
understanding of it can be most readily made by empirical experimentation. 6 .36 .218 6 .31 .429
It is not denied that a theoretical approach could be rewarding; indeed, 12 .33 .161 12 .34 .273
various preliminary studies have been made (Williams, 1976, 1979; Batty and 18 .33 .139 18 .34 .209
Sikdar, 1982). However, the problem is proving to be exceptionally complex 24 .32 .122 24 .34 .172
and it is most easily investigated by empirical means. Furthermore, the 30 .33 .110 30 .34 .144
availability of high-speed computers makes it possible to design aggregation 36 .33 .102 36 .34 .125
experiments of a far more comprehensive nature than would be the case if non- 42 .33 .092 42 .34 .109
automated methods were being employed. Additionally, entire new numerical 48 .33 .082 48 .34 .097
algorithms can be devised to explore different aspects of the aggregation 54 .33 .073 54 .34 .086
problem.
99 .346 .346
The first set of experiments concerns the effects of random aggregation
on the correlation coefficient. Some of the results produced by simple ran- The most interesting discovery here is that scale has no systematic effect
dom aggregation experiments by Gehlke and Biehl (1934) and Blalock (1964) on the mean correlation coefficient. This is because the zoning systems are
have already been described. The question is simply what happens if a more chosen at random so that the sample (or more precisely the zoning) estimates
systematic and comprehensive series of experiments is performed. Interest of the correlation coefficient approximate the population value (which for
is focused on purely random aggregations partly because of the historical zonal data is the observed value prior to the current aggregation, ie the
16 17
99 zone value ). It should also be noted that there is considerable aggregational variability due to the use of simple random zoning systems.
zoning and sampling variability about the mean values but that this reduces An examination of a simple null hypothesis test based on the correlation co-
with increasing numbers of zones or increasing sample sizes. Finally, the efficient shows the sort of additional risk that is involved. For a standard
standard deviations of the zoning distributions are considerably smaller than type I error significance level of 0.05 the value observed from the Monte
the corresponding sampling distributions but exhibit a greater degree of bias. Carlo experiments ranged from .10 to .22, according to the level of spatial
autocorrelation and the particular variable under study.
In these results, somewhere, are the effects of spatial autocorrelation.
Most data sets exhibit positive spatial autocorrelation and the Iowa data Other problems with the sampling-zoning analogy concern the fact that
are no exception. Spatial autocorrelation only affects the zoning distribu- most zonal data sets contain both sampling variability and aggregational vari-
tions because aggregation takes place under contiguity restrictions. Normality ability. In addition, zonal data are unusual in that the population value
or non-normality is not thought to have any important effect on these for any statistic can be determined; for aggregated data this is the value of
experiments. a statistic for the data prior to the current aggregation. A final problem
concerns the fact that geographers have not previously shown any interest in
One way of identifying the effects of spatial autocorrelation is to use studying purely random zoning systems; perhaps they are not thought to be
data sets with different levels of spatial autocorrelation and see what ef- meaningful entities, although it is possible also that until quite recently
fect this has on the zoning distributions. Openshaw and Taylor (1979) de- it was difficult to generate random zoning systems.
scribe a procedure for generating artificial data for the 99 Iowa counties
with the following properties: zero skewness and kurtosis to ensure normality, Another aspect of this discussion concerns the use of inferential statis-
a correlation equal to that observed for the real Iowa data, and regression tical techniques with zonal data. Quite simply, it is seldom clear as to
slope and intercept parameters also equal to the observed Iowa data. Three what is the nature of the hypothesis that is being tested and what, if any-
different levels of spatial autocorrelation were considered (autocorrelation thing, the results signify. If random zoning is not being used then in what
is measured by Moran's I statistic for first order contiguities, see Silk way do zonal data constitute a sample, be it simple or complex? What is the
(1979)): maximum negative spatial autocorrelation, MN, (the best that could population? A statistical answer to some of these questions is to invent a
be achieved were values of -.71 for the vote variable and -.57 for the old 'super population'; for example, that the Iowa data is a random sample of
age variable), zero autocorrelation,Z, and maximum positive autocorrelation, data for Iowa counties because it relates to one, randomly chosen, point in
MP, (the best that could be managed were values of .82 and .92). The same time. While this is easy to say, it is far less easy to identify what the
sets of 10,000 zoning systems as used for Table 7 are applied to these significance tests mean. There is also the difficult problem of determining
artificial data sets with the results shown in Table 8. an appropriate set of sampling error estimation equations. The Iowa data
represents a sample size of 1. Finally, it is not clear as to the geographi-
Table 8. Zoning distributions of the correlation coefficient for cal implications of the hypotheses that could be tested. For example, under
three different levels of spatial autocorrelation what conditions is it possible to compare zonal estimates for one set of
zones with zonal estimates for another set?
number MN Z MP
of standard standard standard (ii) Random aggregation and other statistics
zones mean deviation mean deviation mean deviation
A further consideration is whether or not the results observed for the
6 .31 .443 .61 .294 .60 .247 correlation coefficient also hold good for other statistics. Perhaps the
12 .30 .370 .47 .263 .52 .176 correlation coefficient is a special case. The question is therefore what
18 .29 .350 .42 .227 .48 .142 scale and aggregation variability are likely to be displayed by other un-
24 .31 .309 .40 .192 .44 .121 standardised statistics, such as the mean and the regression slope coefficient.
30 .32 .277 .39 .166 .42 .108 Is it possible that these statistics will be less affected and more robust
36 .32 .242 .38 .146 .40 .098 to aggregation effects? For example, the mean has very good large sample
42 .33 .209 .37 .128 .39 .087 properties. Table 9 should dispel any fears in this direction. It illustrates
48 .33 .183 .36 .112 .38 .080 some results from a regression of percentage rate for Republican candidates
54 .33 .160 .36 .100 .34 .072 as a percentage of the population over 60 years of age (see page 17).

The artificial data with negative spatial autocorrelation has the least
biased results but the largest standard deviations, whereas increasing posi-
tive spatial autocorrelation produces results which are increasingly biased
but with smaller standard deviations. The zero autocorrelation state confers
no particular benefits.
The principal conclusion from these experiments is that the sampling-
zoning analogy does not hold good. There is an additional risk involved in
using standard error formulae for simple random sampling as estimates of the
The mean statistic for the zoning distributions is only very slightly
biased but still has the now characteristic small standard deviation, rela-
tive to the related sampling distributions. The regression coefficient be-
haves in a similar fashion to the correlation coefficient.
(iii) Random aggregation experiments with once aggregated data
The previous experiments concerned the effects of randomly aggregating
zonal data which have already been aggregated at least once previously. Most

18 19
Table 9. Zoning and sampling distributions of a mean and a regression These results are superficially similar to those reported for the re-
slope statistic aggregation of already aggregated data. One difference is the smaller rela-
tive sizes of the standard deviations of the zoning distributions in Table 10.
number mean old aged regression slope This could well reflect the use of small sample sizes; computer times for a
of zoning sampling zoning sampling sample of 100 different individual data aggregations amounted to 2 hours of
zones mean std mean std mean std mean std CPU time on an IBM 370/168. The use of notional contiguities and a sample
data set may also have contributed to reducing the expected range of aggre-
6 14.5 .263 14.5 1.105 1.55 1.071 1.13 1.944 gation effects. Most of the results for the other 50 variables which were
12 14.5 .291 14.5 .747 1.34 .689 1.23 1.088 examined tended to have zoning distribution means of the correlation coef-
18 14.5 .278 14.5 .591 1.27 .569 1.23 .800 ficient which are similar to the 1 km zonal values. Nevertheless, the re-
24 14.5 .267 14.5 .496 1.25 .484 1.24 .649 sults again show that zonal correlations need not correspond to the individual
30 14.5 .249 14.5 .422 1.24 .427 1.24 .536 level correlations and that a 'good' zoning system for one variable can be
36 14.5 .230 14.5 .369 1.23 .389 1.24 .460 quite 'poor' for another, at least in terms of the differences between eco-
42 14.5 .217 14.5 .323 1.23 .346 1.24 .396 logical and individual correlation coefficients. It is still confidently
48 14.5 .202 14.5 .291 1.23 .307 1.24 .352 expected that the aggregational variability in the range of possible results
54 14.5 .186 14.5 .255 1.23 .273 1.25 .312 due to the choice of the first zoning system will exceed that of any subse-
quent re-aggregations of the data, although the current experiment did not
99 14.5 14.5 1.25 1.25 show it. Even if this assumption can be disproven, it is highly likely that
the choice of the first zoning system has a crucial effect on the severity of
Note: std is an abbreviation for standard deviation any subsequent ecological fallacies and that, as far as practicable, the
design of this zoning system should be optimised to minimise these effects.
data that geographers study are of this type. The question arises, therefore, It may be that the possible benefits are slight or are offset by the computer
as to the effects of aggregating data that have not been previously aggre- costs that are involved, but until we try we shall never know.
gated; for example, the aggregation of individual data to a zoning system.
This problem is interesting partly because it is here that ecological falla- (iv) Identifying the limits of the MAUP
cies may be created and because aggregation changes the measurement scale,
usually from a nominal to a continuous form. For example, presence or absence So far attention has been restricted to investigating the variability in
measurements become frequencies or ratios or percentages after aggregation. results due to purely random spatial aggregations. The question now arises
as to what are the worst case or, real limits of aggregation effects if we are
The Sunderland data are used to investigate this problem. The house- perverse enough to look and know bow to find them. The existence of elec-
hold data have 100 metre grid-references attached to them. For this experi- toral boundary gerrymandering has been known about in political geography for
,

ment the 8,483 households can be regarded as single member zones. Notional over 170 years, ever since the famous 1810 gerrymander (Taylor and Johnston,
contiguities can be generated by a Thiessen polygon program so that the in- 1979; pages 371-374). However, it is only recently that its general implica-
-

dividual data zones can be aggregated to form random zoning systems with 25, tions for spatial analysis have beep recognised (Openshaw, 1977a, 1977c,
50, 75, 100, 150, and 200 zones. Table 10 shows the results that were ob- 1978b). By searching for the approximate limits of the range of aggregation
tained for three variables which were selected to show different types of effects it is possible to demonstrate the magnitude and severity of the MAUP.
aggregational behaviour displayed by the correlation coefficient.
Openshaw (1977a) uses a heuristic procedure, of a type similar to itera-
Table 10. Zoning distributions for once aggregated data for Sunderland tive relocation algorithms in cluster analysis, to optimise any general func-
tion by manipulating the zoning systems. This method provides an approxi-
mate solution to what is termed the Automatic Zoning Problem; the algorithm
number variable 1 variable 2 variable 3 is called the Automatic Zoning Procedure (AZP). The basic algorithm is best
of zones mean std mean std mean std described in general terms as consisting of a series of steps.
Step 1. Decide how many regions are required in the final aggregation.
25 .79 .045 -.93 .015 -.94 .014 Step 3 Generate a random zoning system with this number of regions.
50 .82 .034 -.92 .015 -.92 .017 St ep 3.. Randomly select one of these regions and proceed around its bound-
75 .83 .026 -.92 .015 -.91 .016 ary measuring the effects on the objective function of moving
100 .84 .026 -.92 .015 -.90 .020 zones from the bordering regions into it.
150 .83 .022 -.91 .016 -.88 .013 Step 4. 'Once an improvement is recorded for the objective function which
200 .82 .022 -.91 .015 -.87 .018 is being optimised, then check whether the move is possible;
individual that is, it must not destroy the internal contiguity of the
correlation .42 -.81 -.57 region from which a zone is being moved; either reject or accept
the move.
Note: std is an abbreviation for standard deviation Step 5. Once all the members of a region have been examined return to
step 3 to process another region; if all regions have been ex-
amined then go to step 6.
20 21
 Step 6. If one or more moves have been made then return to step 3 other- Yule and Kendall (1950), in a prophetic statement, warn against the
wise stop. development of zonal manipulation procedures of the kind used here. They
In this algorithm the initial data are assumed to relate to a set of zones write 'the student should not now go to the other extreme and claim that,
and these zones are to be aggregated into a smaller number of large zones since a large range of values of correlation coefficients may be obtained
which for purposes of clarity are termed regions. For example, the 99 Iowa according to the choice of a modifiable unit, a particular value has no sig-
Counties form a set of 99 zones which can be aggregated into 6 regions. The nificance' (page 312). Perhaps they did not realise that such a wide range
aggregation is performed in such a way so as to approximately optimise an of aggregation effects were present or did not know how to find them in a
objective function whilst ensuring that all the zones assigned to the same systematic fashion. Instead what they mean is that significance of the cor-
region are internally connected or contiguous. The objective function can relation coefficient depends on the meaningfulness of the areal units on which
be any general function and it need not be continuous. For example, the aim it is based. Perhaps they thought, rather naively, that counties are a sen-
may be to maximise or minimise a correlation coefficient between two vari- sible spatial unit for the study of crop yield relationships whereas arbitrary
ables in order to identify the approximate limits of variability due to the aggregations of the counties to maximise a correlation coefficient would not
MAUP. The AZP algorithm is a heuristic procedure which experience has shown be. It is a shame that Yule and Kendall's work on the modifiable areal unit
can readily solve many types of optimal zoning problems although there is no problem did not continue past this point. Perhaps it could not, because the
guarantee that it will always find the global optimum; indeed with this type problem rapidly becomes one of trying to assess the degree of meaningfulness
of problem there can be no certainty that there is a unique global optimum associated with different geographical definitions for a particular purpose.
to be found. For most problems it probably gets fairly close to a 'good' In general terms it is an impossible problem; for example, how would we go
local optimum; large problems are easier to solve than small ones. No doubt about determining whether counties are an appropriate unit by which to study
the heuristics could be further improved; for example, by the incorporation crop yield relationships or indeed anything?
of a multiple simultaneous move heuristic; but at present this is not the Some critics of the optimal zoning results have suggested that it only
most important problem. More important was the discovery of how to incor- works when applied to correlation coefficients and that in any case the opti-
porate a constraint handling procedure (Openshaw, 1978b), because together mal zoning systems will be of the most peculiar shapes and sizes. This latter
with fast computers this made possible the application of the AZP algorithm point is examined later. The first is simply incorrect. The performance and
to a wide range of region building problems. parameter estimates of a variety of linear and nonlinear models have also been
shown to vary between wide limits (Openshaw, 1977c, 1978a, 1978b).. Some mod-
Returning to the correlation coefficient, this can be used as the ob- els, for instance interaction models, are highly sensitive since the pattern
jective function in the AZP and attempts made to seek zoning systems that of trips that these models try to represent depends on the zoning systems used.
either maximise it or minimise it. This can be regarded as an exercise in A simple example based on the linear regression model should help emphasise
applied gerrymandering or, if you prefer, spatial engineering of zoning sys- the importance of the MAUP. The AZP can be used to produce zoning systems
tems. The dramatic results are shown in Table 11. Even for the 99 Iowa which generate data to either maximise or minimise best statistical estimates
zones, a small data set by current standards, a very wide range of results of the slope coefficient in a regression model based on the Iowa data (Open-
can be obtained. The amount of aggregational variability, or spatial free- shaw, 1978a). In this experiment every time a change is made to the zoning
dom, will be even greater with larger data sets and is probably some expon- system by the AZP the parameters are re-estimated. Two different parameter
ential function of the aggregation factors involved. Nevertheless, for a 6 estimation procedures are used; one based on ordinary least squares the other
region aggregation of the 99 Iowa counties the range of possible correlations on a robust line fitting procedure in the style of Tukey (1977) and described
is between -.99 and +.99. It is also possible that many of the intermediate in McNeil (1977); the purpose is to avoid making normal linear regression
results can be obtained; for example, a zoning system with a correlation of model assumptions. The results are shown in Table 12 and two of the 12 region
0.5 or -0.334. Different amounts of spatial autocorrelation have no notice- zoning systems are shown in Figure 2.
able effects.
Table 12. Approximate limits of regression slope coefficients due to
Table 11. Some approximate limits of the correlation coefficient due to different aggregations of the Iowa data
different aggregations of the Iowa data
number Ordinary least squares estimation robust line fitting estimation
number Iowa data MN data Z data MP data
min r max r min r max r min r max r min r max r of of slope of slope
of zones
zones minimise maximise minimise maximise
6 -.99 .99 -.99 .99 -.99 .99 -.99 .99
.99 -.98 .99 6 -121 27 -84 22
12 -.99 .99 -.97 .99 -.99
.99 12 -24 12 -34 42
18 -.97 .99 -.97 .99 -.97 .99 -.92
.98 18 -12 12 -14 16
24 -.92 .99. -.98 .99 -.90 .99 -.89
.95 24 -8 10 -11 14
30 -.73 .98 -.93 .98 -.86 .98 -.78
.98 -.61 .93 30 -5 7 -12 12
36 -.71 .96 -.93 ..98 -.80
-.52 .93 36 -4 6 -8 10
42 -.55 .95 -.92 .97 -.79 .96
.95 -.39 .89 42 -3 5 -5 8
48 -.50 .90 -.87 .96 -.66
48 -2 4 -4 6
54 -.42 .82 -.85 .95 -.52 .91 -.32 .88
54 -1 4 -2 6
Notes: based on best of five different random zoning systems used as starting
aggregations. MN, Z, MP are the three artificial Iowa data sets 23
(see Table 8)
22
 The propensity that many geographers have shown for attributing substan-
tive interpretations to the slope coefficients in regression models should
be greatly diminished by these results. For example, the value of the slope
coefficient in distance decay models clearly reflects the zoning system as
well as behaviour patterns. It is likely that more complex models, including
entropy maximising spatial interaction models, will also suffer from similar
effects as that displayed by these linear regression models. Currently,
there is no evidence to the contrary.

If the slope coefficients can be made to vary then the performance of

the models can also be made to vary by changing the zoning systems being
studied. This has already been demonstrated by maximising or minimising cor-
relation coefficients. The same effects can be observed for a different good-
ness of fit statistic. Table 13 shows maximum and minimum levels of model
performance as measured by the mean absolute error for the Iowa regression
models.

Table 13. Best and worst fit Iowa regression models

number mean absolute error

of zones worst fit best fit

Figure 2a. Zoning system that minimises the regression slope coefficient 6 14.8 .02
(-24, r = -.25) 12 15.3 .8
18 15.0 .7
24 14.3 1.6
30 12.4 1.9
36 12.2 2.2
42 11.5 2.5
48 10.7 3.2
54 10.3 3.6

Figure 3 shows the geometry of two 12 zone systems that maximise and mini-
mise the mean absolute error. In these experiments the objective function used
in the AZP is the mean absolute error goodness of fit statistic and the model
parameters are re-estimated using a robust line fitting procedure every time
the zoning system changes. The range in results reported here is due solely
to the nature of the zoning systems that are used.

It is now thought likely that no spatial model or method of analysis can

escape the effects of the MAUP. It is also by no means certain that some
methods or models will be better than others in their sensitivity to the MAUP.
It would seem that the range of results tends to be data specific and that it
may be impossible or unwise to try and make any general conclusions other than
the observation that the MAUP is endemic to all spatially aggregated data and
will affect all methods of analysis based upon such data. Its importance
depends on the data and the aggregation factors involved.

(v) Spatial calibration of a statistical model

One interesting, albeit mischievous, development is the use of the AZP

to provide a uniquely geographical approach to estimating the unknown para-
Figure 2b. Zoning system that maximises the regression slope coefficient
meters in statistical models. The conventional approach to estimating the
(12, r = .87)
slope and intercept parameters in a linear regression model is as follows:
24 25
 (1) carefully specify a model; (2) select a 'good' parameter estimation
procedure to yield unbiased estimates of the parameters; and (3) apply the
model to a zonal data set. The choice of the latter whilst not toally hap-
hazard (ie any data set) is usually based on a convenient data set (ie vir-
tually any data set) for an arbitrary set of zones. This procedure is poor
in its geography because of the heresy committed when the data are chosen.
The results could well have a haphazard look about them since no attempt has
been made to control the scale and aggregational variability inherent in the
initial choice of a convenient data set and, in any case, no means are avail-
able for taking these aspects into account.

The analogous purely geographical alternative is to: (1) haphazardly pick

Figure 3 a. Zoning system that produces the worst possible fit
(r = -.058, mean absolute deviation = 15.97, regression
intercept is 60.325 and slope coefficient is -.287)
some convenient values for the undetermined parameters; (2) fit the model by
manipulating the data to fit by optimising the zoning system. The end result
will be similar to the statistical approach except that the initially arbi-
trary parameter values may now have the properties of good estimators. This
geographical approach contradicts the normal science paradigm as it is cur-
rently practised but perhaps this is a necessary violation if we are to es-
cape from the bogus assumption of fixed zonal data. A statistician would re-
gard this geographical approach as unscientific gerrymandering, an exercise
in playing with numbers. But could any geographer possibly recommend the
former statistical approach given its inability to control for the aggrega-
tional variability in zonal data? Both approaches are possible and ideally
some means should be found to combine them.
Consider an example which demonstrates the potential power of the purely
geographical approach. Suppose for the Iowa regression model it is decided
to hold the parameters fixed at some completely arbitrary values; perhaps
geographical theory or prior knowledge could be used to suggest sensible
values. The objective is to fit the model by manipulating the zoning system.

Table 14. Spatial calibration of a linear regression model by seeking opti-

mal 6,12, 18, 24, and 30 zone aggregations of the Iowa data

target parameters zoning systems which zoning systems which pro-

fit a model to these duce data that yield the
intercept slope parameters target parameters

41.46 2.00 12 18 none

41.46 1.75 12 18 24 30 18 24 30
41.46 1.50 6 12 18 24 30 6 12 18 24 30
41.46 1.25 6 12 18 24 30 6 12 18 24 30
41.46 1.00 6 12 18 6 12 18 24 30
41.46 0.75 6 12 18 24 18
41.46 0.50 none none

60.0 1.25 none none

50.0 1.25 12 none
40.0 1.25 6 12 18 24 30 6 12 18 24 30
30.0 1.25 12 18 12 18 24 30
20.0 1.25 12 none
10.0 1.25 none none
Figure 3 b. Zoning system that produces best possible fit
(r = -.997, mean absolute deviation = .322, regression
intercept is -9.054 and slope coefficient is 4.713)

26 27
Figure 4a. Zoning system that fits a model with arbitrary intercept and
slope of 41.4 and 2 (actual 42.4 and 1.90)
Figure 5a. Zoning system that fits model with arbitrary intercept and
slope of 50 and 1.25 (actual 48.4 and 1.26)

Figure 4b. Zoning system that fits a model with arbitrary intercept and Figure 5b. Zoning system that fits model with arbitrary intercept and
slope of 41.4 and 1.0 (actual 42.0 and .98) slope of 30 and 1.25 (actual 30.3 and 1.31)

28 29
Suppose that two sets of runs are performed; the first holds the intercept
at the 99 zone level and systematically varies the slope coefficient; the
second holds the slope coefficient at the 99 zone level and systematically
varies the intercept. An alternative approach to fitting these models is to
minimise the difference between the target parameters and the values esti-
mated for a particular zoning system. Both sets of results are shown in
Table 14 (page 27) with some of the zones being reproduced in Figures 4 and 5
(pp. 28, 29).
The decisions as to whether an acceptable level of fit is achieved are
arbitrary. Nevertheless, it is suggested that quite reasonable levels of
fit have been achieved. It is particularly noticeable that the 99 zone inter-
cept and slope parameters (41.46 and 1.25) can be matched at all five levels
of aggregation and that these zoning systems have zero aggregation effects.
A robust data fitting procedure was used for Table 14. Similar results can
be obtained for ordinary least squares regression, indeed rather more zoning
systems would be judged to fit the target parameters.
One use of spatial calibration is to test specific geographical hypo-
theses about the nature of the results that may be expected; this is elabor-
ated upon later. The argument here is that these empirical results demonstrate
that the statistical and geographical aspects of spatial analysis need to be
integrated. Zone design is in many ways a geographical complement to the
statistical process of parameter estimation and with zonal data they cannot
be separated if meaningful geographical results are to be obtained. This
viewpoint is controversial since it implies that a large number of geographi-
cal studies are: (1) inherently non-geographical, (2) based on haphazard zon-
ing systems with little direct control over aggregation effects; and (3) other-
wise seriously flawed. The logic of this argument leads inextricably to a
very different paradigm for spatial study than that currently used; this is
examined later.
(vi) .. but do the optimal zoning systems look nice?
A final consideration concerns the nature of the optimal zoning systems
shown in Figures 2 to 5. It can be argued, with some justification, that for
reasons not yet investigated or understood, the aggregational properties of
the 'real' zoning systems that geographers use are not as bad as the perverse
optimal zoning systems that the AZP can identify. Perhaps the use of zones
that look 'nice' or are based on regularly shaped units or convenient admin-
istrative definitions may avoid the extremes of the MAUP that have been iden-
tified in the various aggregation experiments. At the limits this is cer-
tainly true but the real problem is that the aggregational properties of
nearly all ad hoc zoning systems are simply unknown. Additionally, it is
difficult to establish any spatial benchmark against which the performance of
alternative zoning systems can be measured. Geometric criteria, shape and
size are not particularly relevant because it is the characteristics of the
data and not the zones themselves that is important. The only absolute
benchmark is the same data at a pre-aggregation or individual level and the
characteristics of the latter are seldom known or available for analysis.
In principle it really does not matter what shape zones have since it
is the relationship between zonal boundaries and the micro-level patterns
which they detect and report that is the subject of spatial analysis. If
the assumption of an isotropic plain were applicable then obviously a
30
geometrically regular set of zones at a carefully selected scale would be most
relevant. However, given the very uneven, lumpy, and discontinuous nature of
real world patterns it is not at all obvious as to why zoning systems should
possess geometric regularity, and if they do what advantages this brings over
the sorts of shapes described in Figures 2 to 5.
Likewise it is not apparent why neutral or locationally arbitrary areal
units, for example grid-squares, should be of any interest in geography.
Since we have the means to design zoning systems which are optimal for a given
purpose, should we not be seeking to use these zoning systems as a means of
investigating further the relationships under study. An analogy with a
television aerial seems most appropriate. You could use an aerial designed
for a radio and perhaps receive a poor picture. You could build your own to
the most beautiful geometric design and get no picture at all. You could
design an aerial to produce the best possible picture without worrying too
much about aesthetics. The zone design problem is broadly analogous to an
aerial. The zoning systems acts as a detector of spatial patterns and the
patterns that are detected and their distinctiveness depend on its design.
Surely no geographer can be content to use zoning systems produced by others
or seek to use nice looking zones purely on aesthetic grounds without any
regard for their performance as pattern detectors.
V POSSIBLE SOLUTIONS
(i) No philosopher's stone
It is not thought likely that a general solution can be found that will
allow existing methods to be used as if the MAUP did not exist. The problem
is far too complex, it is difficult to investigate by analytical means, and
its inherent geographical nature makes it unlikely that a statistical solu-
tion will emerge or if it does that it will suffice.
The simplest solution to the MAUP is to pretend it does not exist and
hope that the results being produced for ad hoc zoning systems will still be
meaningful or least interpretable. This view is implicit, by the lack of
any explicit statements to the contrary, in much geographical work. For
example, the performance of a mathematical model depends partly on its speci-
fication and partly on the zoning system that is used. There is often an
elaborate body of theory to help with the model specification problem but
little or no guidance is available to aid the choice of zoning system.
Likewise many quantitative geography texts describe the existence of the MAUP
but offer little or no advice as how best to use the techniques that are
described to study data for modifiable units.
It is also fortuitous that ad hoc zoning systems often produce plausible
results despite the neglect afforded to the careful definition of areal
entities. However, it should be noted that the general absence of compara-
tive studies may have helped disguise the extent to which zone-dependent
regularities are being uncovered. The principal example sometimes quoted to
demonstrate that the choice of zoning system is of little consequence is that
of factorial ecologies where it seems that the major structural relationships
between sets of social variables are relatively free from zoning effects.
Whilst zonal invariance may be useful for some purposes, it is also slightly
31
worrying that so many social area analyses should be so similar despite cul- statistic is an appropriate measure of the performance of a zoning system.
tural and other important differences. Perhaps a combination of closed number More to the point, how do you decide which criteria to use? How do you know
set problems and correlated denominators have combined to determine the re- if its use is successful?
sults. It may also be that more sensitive methods and carefully engineered
zoning systems would detect very different spatial patterns. An example demonstrates the arbitrariness of these and other general pur-
pose zone design criteria. Table 15 shows the effects on the Iowa correla-
The problem is not that geographers have failed to realise that the MAUP tion coefficient of the following:
exists, only that they do not know what to do about it. Perhaps mistakenly,
they have opted to concentrate on the more tractable statistical problems (i) the equal area, population, and compaction criteria of Sammons (1976);
presented by the analysis of spatial data whilst neglecting the more geo- (ii) the spatial entropy statistic of Batty and Sammons (1978);
graphical ones. The pioneering work on spatial autocorrelation by Cliff and
Ord (1973) and on space-time processes by Bennett (1979) are good examples. (iii) the minimum within-zone heterogeneity criteria of Cliff et al (1975);
They provide elegant solutions to complex statistical problems concerned (iv) the maximum independent variable variance criteria (Cramer, 1964;
with the spatial, and temporal, dependency of zonal data but in so doing they Hannan (1971);
deny the existence of the MAUP. For example, the expected moments of Cliff
and Ord's spatial autocorrelation statistic can be computed under two dif- (v) the maximum relative variation of the independent variable
ferent sets of assumptions, both of which assume that zonal data are fixed. Blalock, 1964);
Yet spatial autocorrelation is a characteristic of zonal data which is de- (vi) the minimum standard error of the regression slope coefficient
pendent on the choice of a particular zoning system. It can be varied by (Williams, 1976).
manipulating the zoning system. All these criteria were formulated as objective functions for the AZP and
solutions obtained.
A final consideration is that when geographers express concern about zon-
ing systems it is mainly a reflection of problems of data comparability. It Table 15. Effects of different zone design criteria on the Iowa correlation
is suggested that the current naive approach depends on two major assumptions coefficient
which are both incorrect. First, that the results will be substantially the
same even if different areal units are used. A corollary of this argument number of zones
would be that meaningful results can be obtained for virtually any set of design criteria 6 12 18 24 30 36 42 48 54
arbitrary areal units; this view is widely held. The aggregation experiments
reported in this section disproves this assumption. Second, that geographers equal area .40 .34 .31 .35 .39 .48 .24 .33 .32
have little or no control over the zoning systems for which data are avail- equal population .88 .72 .63 .56 .59 .47 .50 .40 .55
able so that it is not practical to consider zoning systems as anything other equal density -.03 .71 .52 .52 .53 .53 .53 .46 .56
than fixed. This is an over-simplification because it is always possible to compact zones .30 .12 .25 .30 .46 .03 .42 .26 .21
seek to re-aggregate zonal data in order to find a 'better' set of areal units spatial entropy .90 .21 .26 .28 .54 .26 .33 .43 .46
and thus recover from the effects of the initial aggregation. Why not exploit zonal homogeneity .49 .26 .42 .45 .37 .28 .31 .31 .33
the modifiable nature of areal units rather than passively accept whatever independent variable variation .64 .50 .42 .39 .54 .44 .42 .38 .33
zonal manipulations others perform on their behalf? Sadly, it seems that relative variation .68 .65 .40 .54 .47 .47 .35 .27 .42
many geographers are happier if they do not know about the effects of the standard error of slope .99 .99 .97 .97 .95 .93 .90 .85 .81
zonal manipulations that they or others perform.

(ii) Non-geographical solutions
The most convenient solution is to accept the normal science view that
zoning systems should be independent of the phenomena they are used to re-
port. This would allow the selection of areal units to be independent of the
subsequent analysis, and would partly justify the status quo. However, this
is at best an inherently non-geographical approach. The areal units being
studied should be meaningful in some way which is relevant to the purpose of
the study; therefore, it is argued that zoning systems cannot logically be
independent of the phenomena they represent. In this context independence
implies irrelevance. Nevertheless, a number of arbitrary zone design criteria
have been suggested; for instance, approximate equality of population and
zone shape compaction (Sammons, 1976; 1979); multiple design criteria (Masser
and Brown, 1978); and information statistics (Batty, 1978; Batty and Sammons,
1978). However, it is not apparent why geographers should only be interested
in areal units of a regular shape and size or in what way an information
32
Table 15 demonstrates that different criteria merely produce different
results. At best some of the criteria reduce the systematic effects of scale
but the levels of correlation largely reflect the nature of the criteria.
Since the choice of criteria are arbitrary, then so too are the results.
Worse still, the criteria are independent of any particular purpose so that
the results are largely meaningless.
(iii) A traditional geographical solution
Looked at in another way, the MAUP is fairly trivial. All that is needed
is for geographers to agree upon what constitutes the objects of geographical
enquiry. The MAUP exists because of uncertainty as to what are the spatial
entities which are being studied. Remove that uncertainty and the problem
disappears. Unfortunately, this task of identifying meaningful geographical
entities is a difficult one for many geographers to face because of the
traditional regional geography connotations. Additionally, different defini-
tions will be needed for different purposes.
33
 The best examples of this approach have been the use of functional region
definitions of urban areas for studying census data (Spence et al, 1982;
Coombes et al, 1982). The justification here is that local authority defini-
tions are b st provided by functional region definitions. This solution
e error? If not, then the associated hypothesis must either be rejected or
clearly works well only when there is sufficient geographical knowledge to
define with a high degree of precision the sorts of areal units that are most
sensible for a particular purpose. There are many areas in geography where
it cannot be applied. Furthermore, this approach only removes the aggrega-
tional uncertainty, the effects of the MAUP still survive and condition the
results.
(iv) Towards a new methodology for spatial study
Once it is accepted that the results of studying zonal data depend on the
particular zoning system that is being used, then it is no longer possible
to continue using a normal science paradigm. The data are not fixed so that
the results depend, at least in part, on the areal units that are being
studied; units which are essentially arbitrary and modifiable. The selection
of areal units, or zoning systems, cannot therefore be separate from, or in-
dependent of, the purpose and process of a particular spatial analysis; in-
deed it must be an integral part of it. This view conflicts with the current
use of scientific methods and statistical techniques in geography, and for
these reasons many geographers would refuse to consider it to be a viable
proposition.
Let us continue with the heresy a little longer. The problem is to in-
vent a new paradigm for spatial study which can explicitly handle the geo-
graphy of the MAUP. The most obvious approach is to reverse the normal
science paradigm. Instead of meekly accepting whatever result the choice of
a haphazard zoning systems happens to produce, it is necessary to start by
specifying precisely what outcome is expected. This can take the form of a
hypothesis. If the desired result can be attained by solving the associated
automatic zoning problem, that what are the limits on both the range of re-
sults and the range of zoning systems that produce similar outcomes? What if
anything does the geography of these optimal zoning systems tell us about the
hypothesis being studied? If the desired result cannot be attained without
violating either statistical assumptions or geographical factors, then the
associated hypothesis must be rejected.
The following methodology is suggested as being appropriate for a geo-
graphical solution to the MAUP.
STEP 1. Define the purpose of the study in an explicit fashion. This can be
done by speculating as to what outcome is expected given prior knowledge or
what outcome is desired. For example, does a model that fits data in Nevada
also work in Iowa? This desired result would be expressed as a hypothesis
and set up as an objective function for the AZP. For example, to find out if
a model fits the Iowa data, minimise the model errors using the AZP. If the
question is whether a particular set of parameters can provide an acceptable
l evel of performance then again solve the associated AZP.
STEP 2. Try to obtain the desired result by identifying zoning systems which
approximately optimise the appropriate objective function using the AZP. For
example, minimise the differences between a set of target factor loadings and
values produced for a particular zoning systems; the purpose here might be
to investigate whether a set of social area analysis results for one area
also apply to another.
34
STEP 3. Decide what the results mean in a statistical sense, if this is
appropriate, as well as in terms of the geography of the optimal zoning sys-
tems. Has the target result(s) been achieved with a tolerable degree of
changed, so return to STEP 1. If the results are acceptable, then have any
i mportant statistical assumptions been violated? If constraints are needed
then go to STEP 4. What does the zoning system tell us about the geography
of the study area? The zoning system makes visible the interaction between
the data being aggregated and the hypothesis being studied and a study of
the nature of the zones may be very useful. How did the AZP optimise the
objective function? Is there a trivial spatial solution? If the number of
zones are changed what effect does this have? Finally, are the optimal zoning
systems satisfactory from a geographical point of view?
STEP 4. It may be necessary to introduce constraints to impose restrictions
on either the nature of the zones or on the properties of the data they gen-
erate. These constraints are in addition to the usual contiguity restrictions
necessary to ensure that the zones are internally connected. The AZP can
handle either equality or inequality constraints. These additional constraints
represent a potentially important interface between the geography and statis-
tics of spatial study; for example, constraints to ensure zero spatially auto-
correlated data and a maximum zone size. The feasibility of the additional
constraints is partly related to the aggregation factors involved; when large
numbers of zones are being aggregated then a large number of complex con-
straints can often be satisfied.
STEP 5. Now solve the constrained automatic zoning problem. If a satisfac-
tory result is found then return to STEP 3 for interpretation. If not, then
examine the consequences of failing to satisfy some or all of the constraints.
The extent to which various constraints can or cannot be satisfied may also
provide useful information about the nature of the problem under study. In
most cases an iterative process of experimentation is probably needed.
Clearly the statistical power of this new approach is considerably less
than that promised by conventional methods. Hypothesis testing is used here
as a device for introducing an explicit purpose into the process of spatial
study: This is necessary so that whatever zonal entities are identified
they should be both purpose related and geographically meaningful. The new
paradigm is likely to be most useful when comparative studies are being
performed.
Consider the previous correlation analysis for Iowa. Simply reporting
the level of correlation is not very useful because the result is zone-
dependent. Similarly, it is no use testing the null hypothesis that the cor-
relation coefficient is significantly different from zero. Consider the
related linear regression model. If no prior information other than a model
specification is available, then a wide range of different results can be
obtained depending on the choice of zoning system. However, as we move
through the new paradigm, the introduction of various statistical and geo-
graphical constraints reduces the range of alternatives, although there may
still be a number of different results that require interpretation and
explanation.
For example, suppose we wish to see whether a correlation of 0.8 between
old age and Republican voters reported from some other study area also
35

Figure 6a. Zoning system that minimises the correlation coefficient subject
to constraints (r = -.928, intercept = 5.0, slope = -3.12, spatial
autocorrelation of residuals = 0.0, spatial autocorrelation of
independent variable = 0.0, homoscedasticity = 0.0)
Figure 6b. Zoning system that maximises the correlation coefficient subject
to constraints (r = .993, intercept = -8.598, slope = 4.654,
spatial autocorrelation of residuals = 0.0, spatial autocorrela-
tion of independent variable = 0.0, homoscedasticity = 0.0)
36
occurs in Iowa. Suppose also that you want the associated linear regression
model to satisfy various statistical assumptions; specifically, that the
residuals have zero spatial autocorrelation, that the spatial autocorrelation
of the predictor variable is zero, that the mean residual is zero, and that
the rank correlation between the absolute residuals and the independent vari-
able is also zero (a residual homoscedasticity constraint). The zoning sys-
tems shown in Figure 6 satisfy these assumptions and the range of correlation
is still large, between -.928 and +.993. All we can conclude from this is
that there is so much aggregational variability in these data that the results
are not meaningful despite the undoubted high degree of statistical signifi-
cance. Clearly additional constraints are needed and there should be some
basis for these restrictions. For example, if zonal population sizes are
restricted to about plus or minus 15 per cent of the average, then the range
of correlations is reduced to between .28 and .94. If area is used instead
of population then the range is slightly wider; -.06 to .81. The problem here
is that there is no real basis for these size constraints. The best strategy
would be to combine the Iowa data with another data set ( viz for which the
correlation of 0.8 was obtained). The AZP would be used to identify optimal
zoning systems for both data sets simultaneously (the contiguity constraints
would keep their zones apart). The resulting map patterns for a global
correlation of 0.8 could then be examined.
The idea then is to use the optimal zoning approach to test hypotheses
by manipulating the aggregation process. Instead of asking whether a result
obtained in study area A is different from a result for study area B, it is
necessary to consider the range of results that can be produced for both A
and B. Instead of trying to fit a model to an arbitrary zoning system, it is
necessary to consider which zoning systems provide the best results and to
consider what properties they, or the aggregated data they produce, should
have. The map patterns produced by optimal zoning systems for particular
purposes may themselves contribute to the spatial analysis process.
VI CONCLUSIONS
It has been argued that the MAUP is a fundamental geographical problem
that is endemic to all studies of spatially aggregated data. It is a geo-
graphical fact of life that the results of spatial study will always depend
on the areal units that are being studied. This being so it is time that
geographers started to develop methods of analysis capable of accommodating
and even exploiting this situation. One possible statistical approach could
be based on the role of sampling in statistical inference with the aim of
developing further the obvious similarities between the operations of sampling
and zoning. However, the results reported earlier indicate that the analogy
is a poor one and that there are geographical obstacles to developing it
further; for example, we may have to use random zoning systems. Instead it
has been argued that the MAUP is fundamentally a geographical phenomenon that
is most unlikely to be solved by geographers who are blinkered by both a
statistical perspective and fervent adherence to a paradigm that denies the
very existence of the problem.
This CATMOG has described the first shaky steps in the development of a
new methodology for spatial study which is explicitly based on and around
the purposeful and deliberate engineering of zoning systems. This is viewed
37
as having the potential to open up an entirely new approach to the study of
spatial data as well as offering a general methodological framework into
which any existing model or technique can be incorporated. It is argued that
this constitutes the beginning of a new era which will be characterised by
the development of more relevant and more appropriate core of geographical
analysis techniques. It would seem that the adoption of an exceptionalist
position is a basic prerequisite for this development to take place.
Critics will argue that the 'cure' in the form of AZP appears to be no
better than the disease and will inevitably result in difficulties in making
generalisations outside of a particular zoning system for a particular data
set. The answer to this latter problem is straightforward. All that need be
done is to incorporate several different data sets in the same AZP problem
formulation. They will remain separate entities by virtue of having no con-
tiguity links but they will be linked through the definition of global con-
straints (other than contiguities) and through a common objective function.
The answer to the first point is self-evident. The widespread and serious
impact of the MAUP on spatial study has been convincingly demonstrated so
it is no longer possible to simply ignore it. Thus it would seem that methods
which cannot cope with the MAUP should not be used. Currently there are no
convincing alternative methods for handling spatially grouped data in a
statistically sound framework. So why not investigate more radical non-
statistical frameworks and what could possibly be better for a geographer
than a purely geographical approach?
The consequences of seriously accepting this challenge may well be funda-
mental changes in the manner by which geographers analyse spatial data.
There has to be an admission of an approach to spatial study that is tanta-
mount to operating the normal science paradigm in reverse. This is clearly
non-scientific according to any contemporary liberal definition. It would
seem that while few geographers would question the utility of using the AZP
to identify ranges of possible results due to the MAUP, few have so far shown
any enthusiasm for going any further let alone consider the unimaginable
horrors of scientific heresy. However, it is likely that the former will
inexorably lead to the latter. There have been paradigm shifts before in
science so why not a new one designed specially for geographers? It should
be appreciated that the AZP and its associated methodology offers as yet the
only practical working solution to the MAUP. There can be no real doubts
about its geographical nature but perhaps it is too geographical for many
modern geographers.
It is suggested therefore that the prospect is gradually dawning that the
MAUP is not so much an insoluble problem but rather a powerful analytical
tool ideally suited for probing the structure of areal data sets. The growing
speed of computers opens up the tremendous potential offered by heuristic
solution procedures, such as the AZP, to identify the most appropriate zoning
systems for any particular purpose without having to solve currently intract-
able theoretical and analytical problems. That is to say, we do not as yet
fully understand the problem and we are certainly no way near to being able
to develop a calculus to handle it, but the problem can be solved or turned
around using what are essentially Monte Carlo optimisation methods. Cur-
rently much can be done with small data sets and fairly complex models or with
larger data sets and simple models. Very soon it will be possible to rou-
tinely apply the same methods to any spatial data set and any model or func-
tion, no matter how complex. When this happens often enough then a new geo-
graphical revolution will surely have occurred.
38
BIBLIOGRAPHY
Batty, M. (1978, Speculations on an information theoretic approach to spatial
representation. in: Spatial representation and spatial
interaction, (eds) I. Masser and P.J.B. Brown, (Martinus Nijhoff;
Leiden), pp 115-147.
Batty, M. and Sammons, R. (1978), On searching for the most informative spatial
pattern. Environment and Planning A, 10, pp 747-749.
Batty, M. and Sikdar, P.K. (1982), Spatial aggregation in gravity models.
1. An information-theoretic framework. Environment and
Planning A, 14, pp 377-405.
Bianchi, G., Openshaw, S., Scattoni, P., Sforzi F. and Wymer, C. (1981),
Analisi dell'area sociale: comparazione delle classificazioni con-
dotte su dati medi per sezioni di censimento e su data individuali.
(Paper presented at 2nd Italian Regional Science Conference,
Napoli, October 19th-21st.
Bennett,R.J. (1979), Spatial time series: analysis, forecasting and
control, (Pion: London).
Blalock, H. (1964), Causal inferences in nonexperimental research,
(University of North Carolina Press: Chapel Hill).
Borgatta, E.F. and Jackson, D.J. (1980), Aggregate data: analysis and
interpretation, (Sage Publications: Beverly Hills).
Chapman, G.P. (1977), Human and environmental systems: a geographer's
appraisal, (Academic Press; New York).
Cliff, A.D. and Ord, J.K. (1973), Spatial autocorrelation, (Pion: London)
Cliff, A.D. and Ord, J.K. (1975), Model building and the analysis of spatial
pattern in human geography, Journal of the Royal Statistical
Society Ser B, 37, pp 297-348.
Cliff, A.D., Haggett, P., Ord, J.K., Bassett, K. and Davies, R. (1975),
Elements of spatial structure: a quantitative approach,
(Cambridge University Press: London).
Coombes, M.G., Dixon, J.S., Goddard, J.B., Openshaw, S. and Taylor, P.J.
(1982), Functional regions for the population census of Great
Britain, in: Geography and the Urban Environment, (eds)
D.T. Herbert and R.J. Johnston, (Wiley: London), 5, pp 63-112.
Coombes, M.G. and Openshaw, S. (1982), The use and definition of travel to
work areas in Great Britain: some comments, Regional Studies,
16, pp 141-149.
Cramer, J.S. (1964), Efficient grouping, regression, and correlation in Engel
curve analysis. Journal of the American Statistical
Association, 59, pp 233-250.
Evans, I.S. (1981), Census data handling. in: Quantitative Geography:
a British View, (eds) N. Wrigley and R.J. Bennett, (Routledge
and Kegan Paul: London), pp 46-59.
Gehlke, C.E. and Biehl, H. (1934), Certain effects of grouping upon the size
of the correlation coefficient in census tract material. Journal
of the American Statistical Association, Supplement,
29, pp 169-170.
39
Griffith, D.A. (1980), Towards a theory of spatial statistics, Silk, J. (1979), Statistical concepts in geography, (Allen and Unwin:
Geographical Analysis, 12, pp 325-339. London).
Hannan, M.T. (1971), Aggregation and disaggregation in sociology, Spence, N., Gillespie, A., Goddard, J.B., Kennett, S. Pinch, S. and Williams A.
(Lexington Books: Lexington Mass.). (1982), British cities: an analysis of urban change,
Johnston, R.D. and Rossiter, D.J. (1982), Constituency building, political (Pergamon: Oxford)
representation and electoral bias in urban England. in: Geo- Taylor, P.J. (1977), Quantitative methods in geography, (Houghton
graphy and the Urban Environment, (eds) D.T. Herbert and Mifflin: Boston).
R.J. Johnston, (Wiley: London), 5, pp 113-156.
Taylor, P.J. and Johnston, R.J. (1979), Geography of elections,
Keans, M. (1975), The size of the region-building problem. Environment and (Penguin: Harmondsworth).
Planning A, 7, pp 575-577.
Tukey, J.W. (1977), Exploratory data analysis, (Addison-Wesley:
Masser, I. and Brown, P.J.B. (1978), Spatial representation and spatial Reading, Mass.).
interaction, (Martinus Nijhoff: Leiden).
Williams, I.N. (1976), Optimistic theory validation from spatially grouped
McNeil, D.R. (1977), Interactive data analysis, (Wiley: London). regression: theoretical aspects. Transactions of the Martin
Openshaw, S. (1977a), A geographical solution to scale and aggregation prob- Centre, 1, pp 113-145.
lems in region-building, partitioning, and spatial modelling.
Williams, I.N. (1979), Some implications of the use of spatially grouped
Transactions of the Institute of British Geographers, data. in: Towards the dynamic analysis of spatial systems
New series, 2, pp 459-472.
(eds) R.L. Martin, R.J. Bennett, and N.J. Thrift, (Pion: London),
Openshaw, S. (1977b), Algorithm 3: a procedure to generate pseudo-random ag- pp 53-64.
gregations of N zones into M zones, where M is less than N'.
Environment and Planning A, 9, pp 1423-1428. Yule, G.U. and Kendall, M.G. (1950), An introduction to the theory
of statistics, (Griffin: London).
Openshaw, S. (1977c), Optimal zoning systems for spatial interaction models.
Environment and Planning A, 9, pp 169-184.
Openshaw, S. (1978a), An empirical study of some zone design criteria.
Environment and Planning A, 10, pp 781-794.
Openshaw, S. (1978b), An optimal zoning approach to the study of spatially
aggregated data. in: Spatial representation and spatial
interaction, (eds) I. Masser and P.J.B. Brown, (Martinus Nijhoff:
Leiden).
Openshaw, S. (1981), Le problem de l'aggregation spatiale en geographie.
L'espace Geographique, 1, pp 15-24.
Openshaw, S. (1983), Ecological fallacies and the analysis of areal census
data. Environment and Planning A, (forthcoming)
Openshaw, S. and Taylor, P.J. (1979), A million or so correlation coefficients:
three experiments on the modifiable areal unit problem. in:
Statistical methods in the spatial sciences, (ed) N. Wrigley,
(Pion: London), pp 127-144.
Openshaw, S. and Taylor, P.J. (1981), The modifiable areal unit problem. in:
Quantitative geography: a British View, (eds) N. Wrigley
and R.J. Bennett, (Routledge and Kegan Paul: London), pp 60-70.
Robinson, A.H. (1950), Ecological correlation and the behaviour of individuals.
American Sociological Review, 15, pp 351-357.
Sammons, R. (1976), zoning systems for spatial models, (Redding Geo-
graphical Paper 52, Department of Geography: Reading University).

Sammons, R. (1979), Zone definition in spatial modelling. in: Resources and

Planning, (eds) B. Goodall and A. Kirby, (Pergamon: Ox-
ford), pp 77-100

40